Predicting Three-Dimensional Dose Distribution of Prostate Volumetric Modulated Arc Therapy Using Deep Learning

Area of Science:

• Radiation Oncology
• Medical Physics
• Artificial Intelligence in Healthcare

Background:

• Volumetric modulated arc therapy (VMAT) planning is a complex and time-consuming process in radiation therapy.
• Deep learning offers a method to predict 3D dose distributions, potentially reducing planning time and iterations.
• Accurate dose prediction can maintain treatment plan quality while improving efficiency.

Purpose of the Study:

• To develop, evaluate, and compare three deep learning models for predicting VMAT dose distributions in prostate cancer.
• To assess the impact of different input data structures on model performance.
• To determine the feasibility of using AI for accelerating VMAT treatment planning.

Main Methods:

• Three generative adversarial network (GAN) models were trained using different input data: patient CT alone (PCT alone), patient CT with generalized organ structure (PCTGOS), and patient CT with specific organ structure (PCTSOS).
• Models were trained slice-by-slice on 46 VMAT prostate cancer plans and evaluated on 8 independent plans.
• Prediction time and accuracy metrics, including 3D gamma passing rate and dose-volume histogram (DVH) parameter differences, were analyzed.

Main Results:

• VMAT dose distribution prediction was achieved in approximately 3.5 seconds per patient using the PCTGOS model.
• The PCTGOS model yielded the highest average 3D gamma passing rate (80.51 ± 5.94) and the lowest DVH parameter difference (6.01 ± 5.44%).
• Despite longer prediction times (17.5 s), the PCTSOS model proved most reliable for evaluating multiple dose parameters.

Conclusions:

• AI-driven dose prediction models can significantly accelerate the VMAT treatment planning process.
• These models guide radiation oncologists by providing achievable dose distributions, reducing iterative optimization.
• The PCTGOS model shows promise for efficient and accurate VMAT dose prediction in prostate cancer treatment.